Grouped top-down reasoning with hierarchical window transformer for visual grounding

doi:10.1016/j.ipm.2025.104222

Grouped top-down reasoning with hierarchical window transformer for visual grounding

2025-06-05

0

PRE

AI

DOI：10.1016/j.ipm.2025.104222

Original

Original request for help

Save

Abstract

En 中文

• We propose GTD-HWT, a variant transformer incorporating multi-scale inputs that balances computational efficiency and model performance in visual grounding. • We design a hierarchical reconstruction strategy and dual attention mechanism for efficient cross-scale semantic reasoning from coarse to fine-grained features. • Extensive experiments demonstrate our model outperforms existing methods and achieves comparable results to MLLMs under pre-training paradigm.

Grouped top-down reasoning with hierarchical window transformer for visual grounding

Abstract

Journal

Organization